BUILDING A 100X CHEAPER TRACE JUDGE WITH FIREWORKS

Summary from the TLDR AI newsletter

Fireworks and LangChain developed a cost-effective "perceived error" judge using the Qwen-3.5-35B model, capable of detecting user-identified errors in chatbot interactions. Fine-tuning this judge on chat-langchain data resulted in performance meeting or exceeding frontier models at reduced costs.